In response to the challenges regarding accuracy and real-time performance in vehicle classification detection, this study proposes an improved lightweight model for vehicle detection based on YOLOv5s. The objective to achieve a solution that balances high detection accuracy, swift detection, and low power consumption through a system designed on an FPGA within an MPSoC hardware architecture. In order to make the model more suitable for embedded device deployment, This research replaces the backbone network of YOLOv5s with MobileNetv3 Small and incorporates CBAM attention mechanism and Inner-IoU Loss optimization. This modification aim to achieve lightweighting while enhancing detection accuracy and speed. Compared to the original YOLOv5s model, the enhanced model exhibits a 14.8% increase in mAP,,a reduction of 49.7% in parameters, a 40.7% decrease in model volume, and a 48.9% decrease in computational load..On the NVIDIA 3060 platform, the improved detection speed has surged by 48.8%, reaching 82 FPS. Additionally, hardware acceleration using FPGA has been implemented for YOLOv5s. The optimized system achieves a detection frame rate of 45 FPS while maintaining high precision and speed. This system is easily deployable and suits the demands of intelligent transportation systems, fulfilling the need for efficient real-time monitoring.